Research On Keyframe Extraction And Classification Of Human Motion Capture Data

Motion capture technology obtains the joint Angle information and spatial position information generated by the human body in the process of motion through wearable sensor devices,which will generate a large amount of data in the application process,resulting in storage,retrieval and editing difficulties.A variety of keyframe extraction algorithms have been used to solve this problem,but there are shortcomings such as low extraction efficiency,hard to determine the threshold value and large error of motion data reconstruction.Keyframe contains the key information of motion capture data,which has the advantage of less data and more information,but it is rarely used in the study of motion classification,and its feasibility remains to be verified by experiments.In order to solve the above problems,this paper proposes a new method of keyframe extraction based on linear fitting.The keyframes are extracted by linear piecewise fitting of motion data.Meanwhile,a new motion data reconstruction model is constructed by using linear fitting parameters.An improved residual network model for motion classification using keyframes is proposed.This paper studies from the following three aspects:(1)Keyframe extraction algorithm based on linear fittingAiming at the problem of low extraction efficiency and difficult threshold determination,this paper proposes a keyframe extraction algorithm of motion capture data based on linear fitting.Firstly,gaussian filtering,principal component analysis and standard deviation standardization were used to preprocess the motion capture data.Then the piecewise points are extracted from the motion curve by linear fitting method.Finally,the density clustering algorithm is used to obtain the keyframes from the piecewise point set.The algorithm improves the extraction speed of keyframes,and the extraction rate can reach 100 frames per second.The average compression rate is less than 8%,which effectively reduces the storage of motion data.The reconstruction error is reduced by 22.2%.In this paper,we analyze the interaction between the segmentation determination coefficient R~2 and the neighborhood radius and the number ofεneighborhood samples MinPts,two important parameters of the density clustering algorithm.The results show that the optimal range of R~2is[0.75,0.85],the optimal value ofεis 1,and the optimal range of MinPts is[1,18].(2)Motion data reconstruction technique based on linear fitting parametersAiming at the problem of large error in motion data reconstruction,a motion data reconstruction model based on linear fitting parameters is proposed in this paper.In this model,the by-product linear fitting parameters of keyframe extraction and keyframe index are used for motion reconstruction.The reconstruction results are highly overlapped with the original data,and the error is small,but the data at the keyframe is not smooth.In this paper,a weighted mean filter is designed to solve the problem of uneven results of motion reconstruction,and further reduce the reconstruction error.Compared with traditional reconstruction methods,the reconstruction error of the proposed method is reduced by 44.7%,and the reconstruction speed is 801.5 frames per second.(3)Motion classification model based on improved residual networkAiming at the uncertain feasibility of motion classification based on keyframe,this paper proposes an improved residual network model which can effectively identify motion types using keyframe data.The model is composed of residual network module,bidirectional cyclic network and full link layer,which can classify motion by keyframe.Compared with other sports classification techniques,the accuracy of the model is 99.6%,which is superior to other sports classification techniques and has strong competitiveness.